Telegram voice messages are transcribed and reach the agent as text

[danielehrhardt]

What changed

Voice messages (and round video messages) sent to a Telegram bot used to reach the agent only as an attached .ogg file. Now Godmode transcribes them first and the words become the message. The recording stays attached.

• Key: the OpenAI or ElevenLabs voice key, otherwise an OpenAI or ElevenLabs API tool the chat's agent may use. A tool's key only goes to that provider's own address. The language is detected.
• No key or a provider error: the turn still starts. The agent gets the recording and a note with a neutral reason; the provider's own words stay in the log.
• Answers to questions: the owner can answer an agent's question by voice ("Yes."). The chat echoes what was heard. An answer that couldn't be understood leaves the question open and asks for a typed one.
• Slash commands: spoken words never count as a slash command.
• Typing indicator: shown while the recording is being transcribed.

Tests

• voice.test.ts: key order, API tool keys and their scope, exact host match, empty and oversized recordings, provider errors.
• messaging.test.ts: voice message without a key, with a key, silent recording with a caption, spoken slash, provider refusal, spoken approval and an unheard answer.

Slack and Teams are unchanged.